diff --git a/packager/react-packager/src/lib/GlobalTransformCache.js b/packager/react-packager/src/lib/GlobalTransformCache.js
index d6de72cc3..39bf86e18 100644
--- a/packager/react-packager/src/lib/GlobalTransformCache.js
+++ b/packager/react-packager/src/lib/GlobalTransformCache.js
@@ -11,8 +11,8 @@
 
 'use strict';
 
-const debounce = require('lodash/debounce');
 const imurmurhash = require('imurmurhash');
+const invariant = require('invariant');
 const jsonStableStringify = require('json-stable-stringify');
 const path = require('path');
 const request = require('request');
@@ -21,9 +21,6 @@ const toFixedHex = require('./toFixedHex');
 import type {Options as TransformOptions} from '../JSTransformer/worker/worker';
 import type {CachedResult} from './TransformCache';
 
-const SINGLE_REQUEST_MAX_KEYS = 100;
-const AGGREGATION_DELAY_MS = 100;
-
 type FetchResultURIs = (
   keys: Array<string>,
   callback: (error?: Error, results?: Map<string, string>) => void,
@@ -39,6 +36,94 @@ type FetchProps = {
 type FetchCallback = (error?: Error, resultURI?: ?CachedResult) => mixed;
 type FetchURICallback = (error?: Error, resultURI?: ?string) => mixed;
 
+type ProcessBatch<TItem, TResult> = (
+  batch: Array<TItem>,
+  callback: (error?: Error, orderedResults?: Array<TResult>) => mixed,
+) => mixed;
+type BatchProcessorOptions = {
+  maximumDelayMs: number,
+  maximumItems: number,
+  concurrency: number,
+};
+
+/**
+ * We batch keys together trying to make a smaller amount of queries. For that
+ * we wait a small moment before starting to fetch. We limit also the number of
+ * keys we try to fetch at once, so if we already have that many keys pending,
+ * we can start fetching right away.
+ */
+class BatchProcessor<TItem, TResult> {
+
+  _options: BatchProcessorOptions;
+  _processBatch: ProcessBatch<TItem, TResult>;
+  _queue: Array<{
+    item: TItem,
+    callback: (error?: Error, result?: TResult) => mixed,
+  }>;
+  _timeoutHandle: ?number;
+  _currentProcessCount: number;
+
+  constructor(
+    options: BatchProcessorOptions,
+    processBatch: ProcessBatch<TItem, TResult>,
+  ) {
+    this._options = options;
+    this._processBatch = processBatch;
+    this._queue = [];
+    this._timeoutHandle = null;
+    this._currentProcessCount = 0;
+    (this: any)._processQueue = this._processQueue.bind(this);
+  }
+
+  _processQueue() {
+    this._timeoutHandle = null;
+    while (
+      this._queue.length > 0 &&
+      this._currentProcessCount < this._options.concurrency
+    ) {
+      this._currentProcessCount++;
+      const jobs = this._queue.splice(0, this._options.maximumItems);
+      const items = jobs.map(job => job.item);
+      this._processBatch(items, (error, results) => {
+        invariant(
+          results == null || results.length === items.length,
+          'Not enough results returned.',
+        );
+        for (let i = 0; i < items.length; ++i) {
+          jobs[i].callback(error, results && results[i]);
+        }
+        this._currentProcessCount--;
+        this._processQueueOnceReady();
+      });
+    }
+  }
+
+  _processQueueOnceReady() {
+    if (this._queue.length >= this._options.maximumItems) {
+      clearTimeout(this._timeoutHandle);
+      process.nextTick(this._processQueue);
+      return;
+    }
+    if (this._timeoutHandle == null) {
+      this._timeoutHandle = setTimeout(
+        this._processQueue,
+        this._options.maximumDelayMs,
+      );
+    }
+  }
+
+  queue(
+    item: TItem,
+    callback: (error?: Error, result?: TResult) => mixed,
+  ) {
+    this._queue.push({item, callback});
+    this._processQueueOnceReady();
+  }
+
+}
+
+type URI = string;
+
 /**
  * We aggregate the requests to do a single request for many keys. It also
  * ensures we do a single request at a time to avoid pressuring the I/O.
@@ -46,48 +131,29 @@ type FetchURICallback = (error?: Error, resultURI?: ?string) => mixed;
 class KeyURIFetcher {
 
   _fetchResultURIs: FetchResultURIs;
-  _pendingQueries: Array<{key: string, callback: FetchURICallback}>;
-  _isProcessing: boolean;
-  _processQueriesDebounced: () => void;
-  _processQueries: () => void;
+  _batchProcessor: BatchProcessor<string, ?URI>;
 
-  /**
-   * Fetch the pending keys right now, if any and if we're not already doing
-   * so in parallel. At the end of the fetch, we trigger a new batch fetching
-   * recursively.
-   */
-  _processQueries() {
-    const {_pendingQueries} = this;
-    if (_pendingQueries.length === 0 || this._isProcessing) {
-      return;
-    }
-    this._isProcessing = true;
-    const queries = _pendingQueries.splice(0, SINGLE_REQUEST_MAX_KEYS);
-    const keys = queries.map(query => query.key);
-    this._fetchResultURIs(keys, (error, results) => {
-      queries.forEach(query => {
-        query.callback(error, results && results.get(query.key));
-      });
-      this._isProcessing = false;
-      process.nextTick(this._processQueries);
+  _processKeys(
+    keys: Array<string>,
+    callback: (error?: Error, keyURIs: Array<?URI>) => mixed,
+  ) {
+    this._fetchResultURIs(keys, (error, URIsByKey) => {
+      const URIs = keys.map(key => URIsByKey && URIsByKey.get(key));
+      callback(error, URIs);
     });
   }
 
-  /**
-   * Enqueue the fetching of a particular key.
-   */
   fetch(key: string, callback: FetchURICallback) {
-    this._pendingQueries.push({key, callback});
-    this._processQueriesDebounced();
+    this._batchProcessor.queue(key, callback);
   }
 
   constructor(fetchResultURIs: FetchResultURIs) {
     this._fetchResultURIs = fetchResultURIs;
-    this._pendingQueries = [];
-    this._isProcessing = false;
-    this._processQueries = this._processQueries.bind(this);
-    this._processQueriesDebounced =
-      debounce(this._processQueries, AGGREGATION_DELAY_MS);
+    this._batchProcessor = new BatchProcessor({
+      maximumDelayMs: 10,
+      maximumItems: 500,
+      concurrency: 25,
+    }, this._processKeys.bind(this));
   }
 
 }